Research map / Label-efficient and robust learning / Domain adaptation and generalization

Robust Domain Generalization under Divergent Marginal and Conditional Distributions

NeurIPS 2026 · Label-efficient and robust learning › Domain adaptation and generalization · approach: Foundation-model domain generalization · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes RC-Align, a first-order MAML-based meta-learning method with a domain-class distribution alignment loss and a risk bound decomposing DG error into prior and feature shifts, evaluated on standard DG and multi-domain long-tailed benchmarks.

Problem

Domain generalization typically assumes only P(X|Y) shifts, but real multi-domain settings also have divergent label marginals P(Y), leaving unseen-domain risk under compound shifts poorly controlled.

Approach

Decomposes the query-support risk gap into a prior-shift term (weighted by class-wise support risk) and a feature-shift term (1-Wasserstein distance), shows a domain-class alignment (InfoNCE-style) loss upper-bounds the feature term, and optimizes CE plus this alignment loss inside a FO-MAML LODO loop with manifold mixup.

Key results

• Average DG accuracy 73.0% across PACS, VLCS, OfficeHome, TerraIncognita, versus 72.6% for PCL (best baseline in the table).
• On VLCS-MLT, average 79.1 vs 78.2 for the best BoDA variant; worst-domain 57.4 vs 55.4.
• Ablation on VLCS: ERM 77.5 average, RC-Align full 81.0 average; reported DA loss vs generalization gap correlation r=0.694.

Limitations

• Theory relies on a PL condition and smoothness assumptions that may not hold for deep networks.
• Centroids are estimated from mini-batches, which is noisy for tail classes.

Experimental setup

Datasets: PACS VLCS OfficeHome TerraIncognita PACS-MLT VLCS-MLT OfficeHome-MLT TerraInc-MLT DomainNet-MLT

Benchmarks: PACS VLCS OfficeHome TerraIncognita PACS-MLT VLCS-MLT OfficeHome-MLT TerraInc-MLT DomainNet-MLT

Metrics: average accuracy worst-domain accuracy

Baselines: ERM MLDG CORAL SagNet GMDG SWAD PCL BoDA CRT BSoftmax

Builds on: MAML MLDG Manifold Mixup DomainBed

Closest papers by meaning

• Rethinking Multi-domain Generalization with A General Learning Objective (CVPR 2024 · similarity 0.91)
• Set Valued Predictions For Robust Domain Generalization (ICML 2025 · similarity 0.91)
• Understanding Domain Generalization: A Noise Robustness Perspective (ICLR 2024 · similarity 0.89)
• Unknown Domain Inconsistency Minimization for Domain Generalization (ICLR 2024 · similarity 0.88)
• Sharpness-Aware Model-Agnostic Long-Tailed Domain Generalization (AAAI 2024 · similarity 0.88)
• Towards Generalizing to Unseen Domains with Few Labels (CVPR 2024 · similarity 0.88)